There are significant changes to GPU memory requirements in the latest cryoSPARC release version 0.3.9.
In this new version, a 12GB GPU card can handle refinements at up to 768^3 box size (speed is sacrificed a bit above 640^3 boxes).
The new GPU memory minimum requirement (in bytes) scales as a function of box size like:
20 * box_size ^ 3
So for a 1000 voxel box, this would be about 19GB GPU RAM, and for a 1200 voxel box around 33GB. There are ways we may be able to investigate in the future to reduce this requirement.
For those looking to work with large box sizes:
You’ll need a lot of CPU memory, but ~256GB is probably enough. You may need to go to some different GPU cards. We generally recommend the Titan X which has 12GB of memory and is the best option with that much memory considering performance/$. However with very large volumes, you may need to consider the Teslas which have up to 16GB or, perhaps even better, the Quadro cards which can have up to 24GB of memory (although they cost ~4x the Titan X). We’ve tested with some Teslas, but never with a Quadro although some other users may have.